Support 45t fp8 8 GPU (#3659)

custom_ops/gpu_ops/per_token_quant_fp8.cu

@@ -31,7 +31,8 @@ __global__ void quant_per_token_per_block(const T *input,
     const int num_warp = blockDim.x / 32;
     static constexpr int NUM_PER_THREADS = 128 / 32; // 4
     static constexpr float MAX_VALUE = 448.f;
-    const int end_iter = hidden_size / 128; // warp_iter_num
+    // Note(ZKK) use ceil_div!!
+    const int end_iter = (hidden_size + 127) / 128; // warp_iter_num
     AlignedVector<T, NUM_PER_THREADS> load_vec;
     AlignedVector<float, NUM_PER_THREADS> load_vec_float;
     AlignedVector<phi::dtype::float8_e4m3fn, NUM_PER_THREADS> res_vec;
@@ -42,7 +43,16 @@ __global__ void quant_per_token_per_block(const T *input,
         // deal a block per warp
         for (int iter = warp_id; iter < end_iter; iter += num_warp) {
             const int start_offset = iter * 128;
+
+
+            const bool is_valid_data = start_offset + lane_id * NUM_PER_THREADS < hidden_size;
+
+            if (is_valid_data) {
                 Load<T, NUM_PER_THREADS>(input_now + start_offset + lane_id * NUM_PER_THREADS, &load_vec);
+            } else {
+                #pragma unroll
+                for (int vid = 0; vid < NUM_PER_THREADS; vid++) load_vec[vid] = T(0.f);
+            }
             // get max value per thread
             float max_value_thread = -5e4;
 #pragma unroll
@@ -71,6 +81,7 @@ __global__ void quant_per_token_per_block(const T *input,
                 res_vec[vid] = static_cast<phi::dtype::float8_e4m3fn>(load_vec_float[vid] * MAX_VALUE / max_value_thread);
             }
             // store
+            if (is_valid_data)
             Store<phi::dtype::float8_e4m3fn, NUM_PER_THREADS>(res_vec, quanted_res_now + start_offset + lane_id * NUM_PER_THREADS);
             if (lane_id == 0) {
                 quanted_scale_now[iter] = scale_to_store;
@@ -84,7 +95,10 @@ std::vector<paddle::Tensor> PerTokenQuant(paddle::Tensor& input,
     auto input_dim = input.dims();
     const int token_num = input_dim[0];
     const int hidden_size = input_dim[1];
-    const int hidden_size_scale = hidden_size / block_size;
+    // Note(ZKK) here we use ceil_dive to support 4.5T runing on 8 GPUS
+    // where moe_intermediate_size is 448, can not be divided by 128.
+    const int hidden_size_scale = (hidden_size + block_size - 1) / block_size;
+
     auto quanted_x = GetEmptyTensor(
         {token_num, hidden_size},
         paddle::DataType::FLOAT8_E4M3FN,

fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py

@@ -617,13 +617,13 @@ class BlockWiseFP8MoEMethod(QuantMethodBase):
         ]
         self.up_gate_proj_scale_shape = [
             layer.num_local_experts,
-            layer.moe_intermediate_size * 2 // self.quant_config.weight_block_size[0],
-            layer.hidden_size // self.quant_config.weight_block_size[1],
+            ceil_div(layer.moe_intermediate_size * 2, self.quant_config.weight_block_size[0]),
+            ceil_div(layer.hidden_size, self.quant_config.weight_block_size[1]),
         ]
         self.down_proj_scale_shape = [
             layer.num_local_experts,
-            layer.hidden_size // self.quant_config.weight_block_size[0],
-            layer.moe_intermediate_size // self.quant_config.weight_block_size[1],
+            ceil_div(layer.hidden_size, self.quant_config.weight_block_size[0]),
+            ceil_div(layer.moe_intermediate_size, self.quant_config.weight_block_size[1]),
         ]
         if self.quant_config.is_checkpoint_bf16:
             layer.up_gate_proj_weight = layer.create_parameter(

fastdeploy/model_executor/layers/moe/triton_moe_kernels.py

@@ -14,14 +14,11 @@
 # limitations under the License.
 """
 
+import triton
 import triton.language as tl
 
-from fastdeploy.model_executor.ops.triton_ops.triton_utils_v2 import (
-    paddle_use_triton_v2,
-)
 
-
-@paddle_use_triton_v2()
+@triton.jit()
 def fused_moe_kernel_paddle(
     a_ptr,
     b_ptr,